This work proposes an efficient method for processing 3D data in deep neural networks. The method is evaluated on competitive benchmarks and shows consistent improvements in efficiency while retaining or even improving predictive accuracy. The authors promise to make the code available. Three expert reviewers initially assessed the work as 7/8/6, with minor concerns. The authors provided a detailed rebuttal that was read and discussed by all reviewers. The final assessment is 7/8/7. This work makes a practically useful contribution to applying deep learning on 3D data and the analysis how the speedup is obtained is interesting to a larger audience.